SC Perthub Single Cell Omics

bioRxiv preprint doi: https://doi.org/10.1101/2022.08.20.504663; this version posted January 25, 2023.
The copyright holder for this

preprint (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission.
scPerturb: Harmonized Single-Cell Perturbation Data
Stefan Peidli1,2,x, Tessa D. Green3,x, Ciyue Shen4,5,6, Torsten Gross7, Joseph Min3,
Samuele Garda2,8, Bo Yuan4,5,6, Linus J. Schumacher9, Jake P. Taylor-King7,
Debora S. Marks3,6, Augustin Luna 4,5,6, Nils Blüthgen1,2,+, Chris Sander4,5,6,+
1: Institute of Pathology, Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität

Berlin and Humboldt-Universität zu Berlin, Berlin, Germany
2: IRI Life Sciences, Humboldt-Universität zu Berlin, Berlin, Germany
3: Department of Systems Biology, Harvard Medical School, Boston, MA, USA
4: Departments of Cell Biology and Systems Biology, Harvard Medical School, Boston, MA, USA
5: Department of Data Sciences, Dana-Farber Cancer Institute, Boston, MA, USA
6: Broad Institute, Cambridge, MA, USA
7: Relation Therapeutics, London, UK
8: Institute for Computer Science, Humboldt-Universität zu Berlin, Berlin, Germany
9: Centre for Regenerative Medicine, University of Edinburgh, Edinburgh, UK
x: joint first authors
+: joint senior authors
Abstract
Recent biotechnological advances led to growing numbers of single-cell perturbation
studies, which reveal molecular and phenotypic responses to large numbers of
perturbations. However, analysis across diverse datasets is typically hampered by
differences in format, naming conventions, and data filtering. In order to facilitate
development and benchmarking of computational methods in systems biology, we
collect a set of 44 publicly available single-cell perturbation-response datasets with
molecular readouts, including transcriptomics, proteomics and epigenomics. We
apply uniform pre-processing and quality control pipelines and harmonize feature
annotations. The resulting information resource enables efficient development and
testing of computational analysis methods, and facilitates direct comparison and
integration across datasets. In addition, we introduce E-statistics for perturbation
effect quantification and significance testing, and demonstrate E-distance as a
general distance measure for single cell data. Using these datasets, we illustrate the
application of E-statistics for quantifying perturbation similarity and efficacy. The data
and a package for computing E-statistics is publicly available at scperturb.org. This
work provides an information resource and guide for researchers working with
single-cell perturbation data, highlights conceptual considerations for new
experiments, and makes concrete recommendations for optimal cell counts and read
depth.
1
bioRxiv preprint doi: https://doi.org/10.1101/2022.08.20.504663; this version posted January 25, 2023. The copyright holder for this
Introduction
[Definition of single-cell perturbation data]
Perturbation experiments probe the response of cells or cellular systems to changes
in conditions. These changes traditionally acted equally on all cells in the model
system, such as by modifying temperature or adding drugs. Nowadays, with the
latest functional genomics techniques, single-cell genetic perturbations which act on
individual cellular components have become available. Perturbations using different
technologies target different layers of the hierarchy of protein production (Fig 1). At
the lowest layer, CRISPR-cas9 acts directly on the genome, using indels to induce
frameshift mutations which effectively knock out one or multiple specified genes
(Datlinger et al., 2017; Dixit et al., 2016; Jaitin et al., 2016). Newer CRISPRi and
CRISPRa technologies inhibit or activate transcription respectively (Gilbert et al.,
2014). CRISPR-cas13 acts on the next layer in the hierarchy of protein production to
promote RNA degradation (Wessels et al., 2022). Most small molecule drugs, in
contrast, act directly on protein products like enzymes and receptors and can have
inhibitory or activating effects. When these techniques are applied to large-scale
screens, they create a map between genotype, transcriptome, protein, chromatin
accessibility, and in some cases phenotype (Frangieh et al., 2021). Single cells are
perturbed using unique CRISPR guides, and their corresponding individual barcodes
are read out alongside scRNA-seq, CITE-seq or scATAC-seq reads to identify each
cell’s perturbation condition (Adamson et al., 2016; Dixit et al., 2016; Frangieh et al.,
2021; Rubin et al., 2019). Sequencing with multi-omic readout has been applied to
perturbation experiments only recently. CITE-seq, which assesses surface protein
counts using oligonucleotide-tagged antibodies measured alongside the
transcriptome, has been applied successfully as Perturb-CITE-seq (Frangieh et al.,
2021). Efforts have also been made to link Perturb-seq to an ATAC-seq readout
(Rubin et al., 2019).
[Uses of single-cell perturbation data]

Large-scale single-cell perturbation-response screens enable exploration of complex
cellular behavior not accessible from bulk observation. Directionality in regulatory
network models cannot be inferred without interventional or time-series data about
the system (Gross et al., 2019). Experiments with targeted perturbations can be
modeled as affecting individual nodes of a regulatory network model, while the
molecular readouts provide information on the state changes. This creates the
opportunity to investigate mechanistic processes and infer regulatory interactions
and their directionality (Pratapa et al., 2020). Typically, however, perturbation
datasets have still been too small to elucidate the complexity of a cellular system,
and thus accurately predictive models of regulatory interactions remain difficult to
infer (Gross and Blüthgen, 2020). This limitation will be reduced as dataset size
continues to increase. More directly, drug screens have been used to suggest
therapeutic interventions by analyzing detailed molecular effects of targeted drugs,
2
and designing new single or combinations of perturbations (Bertin et al., 2022; Franz
et al., 2021; Preuer et al., 2018).
[Motivation for a distance measure for high-dimensional expression profiles]

Reliable analysis of increasingly large perturbation datasets requires statistical tools
powerful and efficient enough to harness both the massive number of cells and
perturbations and the inherently high dimensionality of the data. This high
dimensionality complicates calculation of distances between perturbations, as does
cell-cell variation and data sparsity (Kharchenko, 2021). There is not presently a
convention for the statistical comparison measure used in perturbation studies.
Some studies calculate pseudo-bulk by combining all cells in a given perturbation
(Adamson et al., 2016; Datlinger et al., 2017). This means losing any information
about the variation within each cell type. Studies with mixtures of cell types do the
opposite, developing complex methods for quantifying similarity between
heterogeneous cell populations (Burkhardt et al., 2021; Dann et al., 2022; Gehring et
al., 2020). Ideally, one would perform statistical comparisons to identify similar
perturbations and classify perturbation strength using a multivariate distance
measure between sets of cells. Such a distance measure describes the difference or
similarity between sets of cells treated with distinct perturbations, thus inferring
difference or similarity in terms of mechanism or perturbation target; shared
mechanisms tend to produce similar shifts in molecular profiles (Replogle et al.,
2022; Tian et al., 2021). A number of distance measures for scRNA-seq have been
explored by the single-cell community in recent years, including Wasserstein
distances (Chen et al., 2020), maximum mean discrepancy (Lotfollahi et al., 2020),
neighborhood-based measures (Burkhardt et al., 2021; Dann et al., 2022),
E-distance (Replogle et al., 2022), Here we exclusively use the E-distance, a
fundamental statistical measure of distances between point clouds that can be used
in a statistical test to identify strong or weak perturbations as well as to distinguish
between perturbations affecting distinct cellular sub-processes (see Methods). This
test is a statistically reliable tool for computational diagnostics of information content
for a specific perturbation and can inform design of experiments and data selection
for training models.
[Motivation for unifying datasets]

Each large perturbation screen is specifically designed to study a particular system
under a set of perturbations of interest. This results in a heterogeneous assortment
of single-cell perturbation-response data with a wide range of different cell types,
such as immortalized cell lines and iPSC-derived models, and different perturbation
technologies, including knockouts, activation, interference, base editing, and prime
editing (Przybyla and Gilbert, 2022). Novel computational methods to efficiently
harmonize these different perturbation datasets on a large scale are needed. Such
integrative analysis is complicated by batch effects and biological differences
between primary tissue and cell culture (Forcato et al., 2021; Luecken et al., 2022).
Published computational methods for perturbation data are primarily focused on
3
individual datasets (Duan et al., 2019; Jin et al., 2022; Lotfollahi et al., 2019). Moving
from single-dataset to multi-dataset analysis will require development of principled
quantitative approaches to perturbation biology; the dataset resource based on this
work can serve as a foundation for building these models going forward.
[Prior work]
While large databases of perturbations with bulk readouts exist, single-cell
perturbation technologies are newer and data is still not unified (Stathias et al., 2020;
Tsherniak et al., 2017). Existing collections of datasets are primarily a means for
filtering datasets but do not supply a unified format for perturbations. Some of these
collections of single-cell datasets were produced to benchmark computational
methods for data integration (Lance et al., 2022). Another collection specifically
aggregated all available data in the single-cell literature, with well over a thousand
datasets described in a sortable table, but with no attempt at data unification and
labels focused on observational studies (Svensson et al., 2020). The Broad
Institute’s Single Cell Portal provides .h5 files for 478 studies with cell type names
from a common list but does not harmonize the datasets or allow for filtering by
perturbation (Broad Institute, 2022). Yet, unified datasets are key for developing
generalizable machine learning methods and establishing multimodal data
integration. A recent review and repository of single-cell perturbation data for
machine learning lists 22 datasets but supplied cleaned and format-unified data for
only 6 (Ji et al., 2021). An existing unified framework for single cell data, called
‘sfaira’, is ideal for model building and memory efficient data loading, but the public
‘data zoo’ does not currently supply perturbation datasets or standardized
perturbation annotations (Fischer et al., 2021).
[Our contribution]
We aim to provide a resource of standardized datasets reporting targeted
perturbations with single-cell readouts and to facilitate the development and
benchmarking of computational approaches in system biology. We collected a set of
44 publicly available perturbation-response datasets from 25 papers (Table 1, Supp
Fig 3B). Our perturbation strength quantification and comparison of
perturbation-specific variables, such as the number of perturbations and the number
of cells per perturbation, across experiments may serve as a reference for optimal
experimental design of future single-cell perturbation experiments. We also describe
the E-distance and E-test as tools for statistical comparisons of sets of cells and
benchmark their robustness and applicability for distinguishing both perturbations
and cell types across different datasets and modalities. A web interface for data
access, analysis and visualization is available at scperturb.org, and a Python
implementation of e-distance based statistics for single cell data is publicly available
as scperturb on PyPI.
4
Fig 1: Perturbation-response profiling for single cells. Different perturbations act at

different layers in the hierarchy of gene expression and protein production (purple arrows).
Perturbations included in scPerturb include CRISPR-cas9, which directly perturbs the
genome; CRISPRa, which activates transcription of a target gene; CRISPRi, which blocks
transcription of targeted genes; CRISPR-cas13, which cleaves targeted mRNAs and
promotes their degradation; cytokines that bind cell surface receptors; and small molecules
perturb various cellular mechanisms. Single cell measurements probe the response to
perturbation, also at different layers of gene expression: scATAC-seq directly probes
chromatin state; scRNA-seq measures mRNA; and protein count data currently is typically
obtained via antibodies bound to proteins.
Results
[Overview of datasets in the information resource]
Molecular readouts for our 44 publicly available single-cell perturbation response
datasets include transcriptomes, proteins and epigenomes (Table 1, Fig 2A).
Metadata was harmonized across datasets (Supp Table 2). 32 datasets in this
resource were perturbed using CRISPR and 9 datasets perturbed with drugs. This
paucity of drug datasets is likely due to the experimental hurdle of applying large
numbers of perturbations to cells; although it is possible to set up multiplexed
sequencing for arrayed treatment conditions, this entails a large amount of manual
labor necessary to set up hundreds of separate wells with individual drugs, limiting
the total number of drug perturbations in a single experiment. In contrast, the mixed
set of single guides for CRISPR perturbations can be applied in parallel, allowing
5
these experiments to be scaled up massively. While 32 datasets measure

scRNA-seq exclusively, we also include scATAC from three papers, including one
with simultaneous protein measurements (Mimitou et al., 2019). For each
scRNA-seq dataset we supply count matrices, where each cell has a perturbation
annotation, quality control metrics including gene counts and mitochondrial read
percentage. Quality control plots for each dataset are also available on scperturb.org
(for an example see Supp Fig 1). Three CITE-seq datasets are included with protein
and RNA counts separately downloadable (Frangieh et al., 2021; Papalexi et al.,
2021).
[Choice of features in scATAC-seq data]

In contrast to scRNA-seq data, which can be represented naturally as counts per
gene, there is no single obvious feature that could be computed for scATAC-seq
data, which, in its raw form, provides a noisy and very sparse description of
chromatin accessibility over the entire genome. We therefore generated five different
feature sets for scATAC-seq data, as motivated in prior studies (Chen et al., 2019;
Granja et al., 2021). These either attempt to summarize chromatin accessibility
information over different types of biologically relevant genomic intervals (e.g. gene
neighborhood), or represent dense low-dimensional embeddings of the original data
(see Methods for details). Different features address different biological questions in
different contexts (Pierce et al., 2021), which is why we include all five feature sets in
our resource.
[Count statistics in scRNA-seq]

Sample quality measures vary significantly across datasets (Fig 2B). Total unique
molecular identifier (UMI) counts per cell and number of genes per cell are
calculated as described in (Luecken and Theis, 2019). The average sequencing
depth, i.e. the mean number of reads per cell, in each study affects the number of
lowly expressed genes observed. Increasing the sequencing depth increases the
number of UMI counts measured even for lowly expressed genes, reducing the
uncertainty associated with zero counts (Haque et al., 2017; Svensson et al., 2020).
These differences can affect the distinguishability of perturbations and performance
of downstream analysis methods.
[Cells per perturbation]

The total number of cells per dataset is usually restricted by experimental limitations,
though has increased over time (Supp Fig 2A). Therefore, there is a tradeoff
between the number of perturbations and the mean number of cells per perturbation
in a dataset (Fig 2C). The type of perturbation partially dictates the number of cells
per perturbation; CRISPR datasets tend to have more perturbations than drug
datasets because they are easier to scale up using multiplexing, with a
corresponding smaller number of cells per perturbation. This is visible in Figure 2B,
where CRISPR datasets are clustered to the left of the plot.
6
Fig 2: Single cell perturbation-response datasets are diverse in type, size, and quality.
(A) The majority of included datasets result from CRISPR (DNA cut, inhibition or activation)
perturbations using cell lines derived from various cancers. The studies performed on cells
from primary tissues generally use drug perturbations. Primary tissue refers to samples
taken directly from patients or mice, sometimes with multiple cell types. (B) Sequencing and
cell count metrics across scPerturb perturbation datasets (rows), colored by perturbation
type as in Figure 2C. From left to right: Distribution of total RNA counts per cell (left);
distribution of the number of genes with at least one count in a cell (middle); distribution of
number of cells with at least one count of a gene per gene (right). Most datasets have on
average approximately 3000 genes measured per cell, though some outlier datasets have
significantly sparser coverage of genes. (C) Each circle represents one dataset. Due to
experimental constraints, most datasets have approximately the same number of total cells,
pooled across a set of marked perturbations, resulting in a tradeoff between the number of
perturbations and the number of cells in each perturbation. CRISPR-perturbation datasets,
compared to drug-perturbation datasets, have fewer cells per perturbation but a larger
number of perturbations. Due to experimental constraints, most datasets have approximately
the same number of total cells, pooled across a set of marked perturbations, resulting in a
tradeoff between the number of perturbations and the number of cells in each perturbation.
7
Fig 3: E-statistics describe distinctiveness of perturbations in single-cell data. (A)

Definition of E-distance, relating the width of cell distributions of high-dimensional molecular
profiles to their distance from each other (see Methods). A large E-distance of perturbed
cells from unperturbed indicates a strong change in molecular profile induced by the
perturbation. (B) Distribution of E-distances (plus 1 for log scale, same in Fig 3C) between
perturbed and unperturbed cells across datasets. The number of perturbations per dataset is
displayed along the bottom. Note that this plot is best used to compare the shape of the
E-distance distribution rather than the magnitude; the mean E-distance will vary significantly
with other dataset properties. (C-E) Analysis based on E-statistics for one selected dataset
(Norman et al., 2019): (C) Distribution of E-distances between perturbed and unperturbed
cells as in Figure 3B. Each circled point is a perturbation, i.e., represents a set of cell
profiles. Each perturbation was tested for significant E-distance to unperturbed (E-test). (D)
Pairwise E-distance matrix across the top and bottom 3 perturbations of Figure 3C and the
unperturbed cells. (E) UMAP of single cells of the weakest (left, Bottom 3) and strongest
(right, Top 3) perturbations.
[E-distance: definition]
To compare and evaluate perturbations within each dataset we utilized the
E-distance, a statistical distance measure between two distributions, which was used
as a test statistic in (Replogle et al., 2022). Essentially, the E-distance compares the
mean pairwise distance of cells across two different perturbations to the mean
pairwise distance of cells within the two distributions (see Methods). If the former is
much larger than the latter, the two distributions can be seen as distinct. Similar to
8
Replogle et al., we compute the E-distance after PCA (Principal Component

Analysis) for dimensionality reduction (see Methods).
[E-distance: interpretation]
The E-distance provides intuition about the signal-to-noise ratio in a dataset. For two
groups of cells, it relates the distance between cells across the groups (“signal”), to
the width of each distribution (“noise”) (Fig 3A). If this distance is large, distributions
are distinguishable, and the corresponding perturbation has a strong effect. A low
E-distance indicates that a perturbation did not induce a large shift in expression
profiles, reflecting either technical problems in the experiment, ineffectiveness of the
perturbation, or perturbation resistance.
9
10
Fig 4: Cell type hierarchies computed using E-distance match known cell type
relationships. (A) Hierarchical clustering of pairwise E-distances computed using RNA
matches prior knowledge of transcriptome-defined cell types. Dendrogram and heatmap use
the same distances. Data from (Hao et al., 2021). (B) As in (A) but using antibody-tagged
surface proteins instead of RNA. (C) Visualization of cell type relationships in full multimodal
dataset after batch correction. Coordinates and cell type annotations from (Hao et al., 2021).
[E-distance: distinguishing cell types]

To test whether E-distance values replicate differences between well-known cell
types, we applied the E-distance to a CITE-seq human PBMC (peripheral blood
mononuclear cells) dataset with existing cell type annotations (Hao et al., 2021).
Separately for RNA and protein (from antibody-derived tags), we computed
PCA-based E-distances between all pairs of cell types, equivalent to how
perturbation E-distance is computed. The resulting pairwise distance matrices were
used to compute cell type hierarchies (see Methods), which we compare to known
cell type relationships (Fig 4A, 4B, Diehl et al., 2016). In both data modalities, B cell
subtypes are clustered together, and platelets are the most distinct from any other
cell type. Lymphoid and myeloid cells form two separate groups in the E-distance
hierarchy. Notably, innate lymphoid cells (ILCs) and NK cell clusters form a distinct
group as well. ILCs are innate immune cells that functionally correspond to specific
types of classical lymphocytes properly expressing diversified antigen receptors; NK
cells are a type of ILC also known as ILC4 and are functionally similar to cytotoxic T
cells (Artis and Spits, 2015; Vivier et al., 2018). This functional similarity translates to
strong similarities in transcriptional profiles, which often leads to difficulties in
distinguishing NK cells and cytotoxic T cells in scRNA-seq data. These cell types are
more easily disentangled by protein marker based distances, exemplifying the
usefulness of CITE-seq as a method for identifying immune cell types. Likewise,
when using protein, T cells are clustered primarily by CD4/CD8 type, whereas, when
using RNA, they are clustered by functional phenotype (naive, proliferating,
memory). For instance, clustering the cells with the E-distance in RNA-space
separates proliferating cells of many types into a single cluster, likely due to shared
expression of cell-cycle related genes; the cell cycle is known to have a strong effect
on the transcriptome profile of cells and is not captured by surface protein
measurements. We conclude that, by comparing RNA and protein representations,
the protein modality more accurately represents cell type differences traditionally
defined by immunologists on the basis of surface proteins, whereas the RNA
representation primarily reflects functional programs of the cells such as cytotoxicity
or proliferation. In both cases, the E-distance accurately captures known
characteristics of each measurement modality.
[E-distance: E-test]
The E-distance can also be used as a test statistic to assess whether cells after a
perturbation are significantly different from unperturbed cells (Replogle et al., 2022),
Supp Table 3). The E-test is a permutation test that uses the E-distance as a test
11
statistic (Székely and Rizzo, 2013, details in Methods). This permutation test
requires hundreds of iterations of computing the E-distance on randomized data, but
is necessary for direct comparisons of perturbations across studies, as the exact
value of the E-distance depends on dataset-specific parameters. The lack of
robustness is exemplified by Figure 5A, which indicates that reducing the number of
cells actually increases the E-distance, while the E-test gradually loses significance.
Thus, while the E-distance is a useful tool for analysis within one dataset or between
experimentally similar datasets, we recommend the E-test as the appropriate
statistical measure for comparisons of perturbation effects between different
datasets.
[E-distance: perturbation dataset example]

Interestingly, we found that E-distances between perturbed and unperturbed cells
vary significantly across datasets (Fig 3B). The dataset labeled with
“NormanWeissman2019” had the largest mean E-distance between all perturbations
(Norman et al., 2019) compared to datasets of similar size. In fact, expression
profiles of most perturbations in this dataset were significantly different from those of
unperturbed cells according to the E-test (Fig 3C). Plausibly, this is in part caused by
two-target perturbations using CRISPRa in that dataset: targeting the same gene
with two single guides increases the chances of causing a considerable change in
the transcript profile. Indeed, the three perturbations with highest E-distance are
double perturbations while the three closest in E-distance are not. The
corresponding UMAPs for these perturbations, computed using the same PCs as the
E-distance, provide a confirmatory visual intuition for high and low E-distances (Fig
3E). The top three perturbations causing the largest E-distance to unperturbed are
easily distinguishable from the gray unperturbed cells, while the bottom three
weakest perturbations are part of a single, uniform cloud virtually indistinguishable
from the unperturbed cells. The smallest E-distance thus results from perturbations
which have the least effect on the distribution of cells.
[E-distance: perturbation-perturbation distances]

The E-distance can also be used to measure similarity between different
perturbations. For instance, there is a clear overlap of CEBPE+SPI1 and
JUN+CEBPA perturbed cells in the UMAP (Fig 3E). This overlap is captured by the
low E-distance between the two perturbations; these two perturbations are closer to
each other than they are to unperturbed cells or to other perturbations (Fig 3D, Supp
Fig 3A). We envision that the E-distance can be used as a suitable distance for other
downstream tasks such as drug embeddings and clustering of perturbations, which
could allow inference of functional similarity of perturbations by similarity in their
induced molecular responses measured by the E-distance.
12
Fig 5: Effect of subsampling UMI counts per cell and number of cells per perturbation
on E-statistics. (A) E-distance of each perturbation to unperturbed in Norman et al. while
subsampling the number of cells per perturbation; Color indicates E-test results;
“significance lost”: perturbation significant when all cells are considered, but not significant
after subsampling. The E-test loses significance with lower cell numbers while the
E-distance actually increases. (B) Overall number of perturbations with significant E-test
decreases when subsampling cells. (C) As in Figure 5A but subsampling UMI counts per cell
while keeping the number of cells constant. Loss of E-test significance and dropping
E-distance to unperturbed as overall signal gets deteriorated with removal of UMI counts. (D)
As in Figure 5B but subsampling UMI counts per cell while keeping the number of cells
constant.
[E-distance: Effect of number of cells on the E-statistics]

We investigate the robustness of E-distance and E-test scores to experimental and
computational parameters using our extensive collection of harmonized single-cell
perturbation datasets. We subsampled the number of cells per perturbation to create
artificially smaller datasets, then examined how the E-distance and E-test results
change. We find that the E-distance actually increases as the number of cells per
perturbation decreases, indicating that cells per perturbation should be standardized
prior to calculating E-distances (Fig 5A). This increase reflects the fact that the
13
E-distance, as a V-statistic, is a biased estimator (Székely and Rizzo, 2013). Despite

the increase in E-distance with falling cell numbers, the number of significant
perturbations correctly decreases with fewer cell counts, and only some datasets
have saturated significance at full number of cells in that dataset (Fig 5B). This
saturation point will depend on the strength of the perturbation and on the
heterogeneity of the dataset; if all cells are similar to each other, a small set of cells
will sufficiently describe every possible response to a perturbation. This suggests
that, unsurprisingly, increasing sample size enables discovery of significant
perturbations with smaller magnitude.
[E-distance: Effect of number of UMI counts on the E-statistics]

Similarly, we subset the number of UMI (unique molecular identifier) counts per cell,
finding that E-distance increases as the number of UMI counts per cell increases
(Fig 5C). The number of significant perturbations under the E-test, though, saturates
around 500 counts per cell, with most perturbations that were significant at the full
measured read depth maintaining that significance even with far fewer counts per
cell (Fig 5D). The stability of E-test results with respect to UMI counts, in contrast to
the actual E-distance value, exemplifies the necessity of the E-test as the
appropriate statistical measure to evaluate perturbation effects. The optimal UMI and
cell counts for a given experiment depend on downstream specific modeling tasks,
as discussed in more detail elsewhere (Gross and Blüthgen, 2020). As a baseline for
significant perturbations, as defined by the E-test, we suggest at least 300 cells per
perturbation (Fig 5B) and 1000 average UMI counts per cell (Fig 5D) as an
experimental guideline for distinguishable perturbations.
[E-distance: robust to calculation choices]

We also examined how choices made in computing the principal components (PCs)
used in distance calculations affects E-statistics. The number of PCs used from PCA
to compute the E-distance had a moderate effect on E-test results, mildly decreasing
the number of significant perturbations (Supp Fig 5A). Interestingly, E-test
significance was lost most rapidly in a TAP-seq (targeted perturb-seq) experiment
(Schraivogel et al., 2020). TAP-seq only measures approximately 3000 genes of
interest, and thus has far fewer starting features than other datasets. This leads to
reduced correlation between genes in the resulting expression matrix, and thus
fewer PCs are needed to sufficiently describe the data. The number of highly
variable genes (HVGs) used to compute the PCs had almost no effect on E-testing
above 500 HVGs (Supp Fig 5B). Computing PCs separately for each perturbation
rather than jointly across all perturbations in a given dataset similarly had minimal
impact on the resulting E-distances (Supp Fig 5C). Taken together, this analysis
indicates that E-statistics can be calculated as part of an existing computational
workflow, which already includes calculating PCs.
14
[Dataset highlights]
With this assembly of datasets and quality control metrics described below, we were
able to nominate notable datasets. The most extensive drug dataset is the sci-Plex 3
dataset with over 188 drugs tested across three cell lines (Srivatsan et al., 2020);
107 of those perturbations were significant according to E-test analysis (Supp Table
3). Five drugs in this dataset also appear in other drug perturbation datasets (Supp
Table 4). We hope that future large-scale drug screens will enable more detailed
analysis of drug response across different cell types and conditions. Another drug
dataset applies combinations of three drug perturbations at varying concentrations
across samples (Gehring et al., 2020). We excluded this dataset from the E-distance
analysis due to its complex study design, which was not directly comparable to any
other included studies. By far the most detailed CRISPR dataset is from a recently
published study which perturbed 9867 genes in human cells
(ReplogleWeissman2022). Containing >2.5 million cells, this dataset is the largest in
our database, with the number of cells each gene is detected in significantly higher
than in other datasets (Fig 2B). Notably, 138 CRISPR perturbations are seen in both
RNA and ATAC datasets (Supp Table 5). More than 100 genes perturbed with
CRISPRa in one dataset are perturbed with CRISPRi perturbations in another
dataset of the same cell line, either in one paper (Tian et al., 2021) or across multiple
studies (Norman et al., 2019; Replogle et al., 2022). The most frequently perturbed
gene, MYC, is perturbed in 9 datasets from 3 papers. Protein, RNA and ATAC
readouts for CRISPRi perturbation of MYC are all available for K562 cells (Frangieh
et al., 2021; Pierce et al., 2021; Replogle et al., 2022).
Discussion
[Concise summary of results]
We present a dataset resource and an intuitive analytic method for quantifying and
analyzing single-cell perturbation datasets. Datasets are described in detail, with
additional individualized quality control metrics available on scperturb.org. The
uniform annotations in this resource will enable data integration and benchmarking
as well as exploration of shared perturbations across datasets. The use of the
E-distance is motivated and applied to quantitatively compare perturbations within
each dataset. We illustrate how to interpret high and low E-distances and use
E-distances to identify functionally similar versus distinct perturbations. We also
investigate the effect of dataset specific parameters on E-statistics, showing that
E-statistics stabilize above 1000 counts per cell and 300 cells per perturbation.
[Overlap of perturbations across studies]

While this work simplifies access to datasets, joint analysis of single-cell datasets is
limited by the complexity of data integration. Across the eight drug datasets
examined in this study, only 5 chemical agents occurred in more than one dataset
(Supp Table 4). Shared gene targets are found more often across the CRISPR
datasets (Supp Table 5). However, multiplicity of infections and other conditions
15
frequently differ as well, and comparisons are further complicated by distinct

perturbation methods (Table 1). The considerable overlap of perturbations across
studies makes this a useful resource for benchmarking model generalizability. With
more datasets anticipated in the future, we will have the unique opportunity to
integrate datasets with more overlapping perturbations and nominate machine
learning benchmarks for data integration.
[Towards standardization]
Lack of standardization in data sharing and processing hampered the creation of this
resource. Although many processed datasets were available on the NCBI Gene
Expression Omnibus (GEO) (Barrett et al., 2013), there is no standard format for
sharing CRISPR barcode assignments and other metadata. Starting analysis from
sequencing reads may have improved interoperability of datasets in this resource,
but guide assignment procedures and demultiplexing algorithms are specific to
experimental setup. For scATAC data, data comparison is made more challenging by
the lack of a standard method for feature assignment (see Methods). In particular,
scATAC feature assignments specific to CRISPR perturbations, where known
locus-of-action could be used to improve feature calls (Chen et al., 2019). In all
modalities, many datasets only supplied processed data, or raw data was only
available after institutional clearance. Adding more datasets to this resource, or the
creation of similar resources in the future, would be much easier if there were
standard formats for sharing perturbation data, and, more generally, standard
formats for sharing single-cell annotations. We think a community-wide discussion
on standardization of such data is urgently needed, as has been done for proteomic
data (Gatto et al., 2022).
[Additional considerations for single cell perturbation experimental design]

Experimental design choices such as the optimal number of cells per perturbation
and the sequencing depth for each cell depend on the questions the dataset is
intended to answer, and on the strength and uniqueness of the gene expression
changes caused by the perturbations (Fig 2B). Unfortunately, it is difficult to ascertain
to what extent a low E-distance between perturbed and unperturbed cells in the data
is caused by technical noise. Increasing the dose or varying the time between
perturbation start and harvesting of the cells may be advisable to increase the signal
to noise ratio without sequencing more cells. For perturbation distinguishability as
defined by the E-test, regardless of experimental parameters, we find that one
should have at least 300 cells per perturbation and an average of 1000 UMIs per
cell.
[Conclusion]
We envision that the scPerturb collection of datasets and the suggested E-statistics
analytic framework will be a valuable starting point for the analysis of single-cell
perturbation data. The unified annotations and perturbation significance testing
across datasets should prove especially useful to the machine learning community
16
for training models on this data. We expect new datasets and new experimental
perturbation methods in the future will enable the community to develop novel
computational approaches which exploit the increasing amount and complexity of
single-cell perturbation data, aiming at the development of increasingly accurate and
quantitatively predictive models of cell biological processes and the design of
targeted interventions for investigational or therapeutic purposes.
Data Availability
The website scperturb.org stores harmonized datasets with the following:
● scRNA-seq and antibody-based protein datasets: .h5ad files and .mtx files are
available, which can be easily read with python or R scripts.
● scATAC-seq: multiple different feature matrix definitions as separate download
options.
● Access details for the original publication for each dataset
● Quality control plots for each dataset
● Filtering, e.g., by readout or type of perturbation
● RNA data at https://zenodo.org/record/7041849 and ATAC data at
https://zenodo.org/record/7058382
Code Availability
Open access source code is at https://github.com/sanderlab/scPerturb/. We
compiled a corresponding Python package called scperturb for performing
E-statistics (E-distance and E-testing) in single-cell data, published on PyPI under
https://pypi.org/project/scperturb/.
17
Table 1: Key metadata for datasets on scPerturb.org. More details in Supp Table 1.
Source Paper Modality Perturbation type Number of
perturbations
(Adamson et al., 2016) RNA CRISPRi 9, 20, 114
(Aissa et al., 2021) RNA drugs 4
(Chang et al., 2022) RNA drugs 4
(Datlinger et al., 2017) RNA CRISPR-cas9+TCR& 97
(Datlinger et al., 2021) RNA CRISPR-cas9+TCR& 48
(Dixit et al., 2016) RNA CRISPR-cas9 31
(Frangieh et al., 2021) RNA + protein CRISPR-cas9 249
(Gasperini et al., 2019) RNA CRISPRi 43314*, 39087*, 16531*
(Gehring et al., 2020) RNA drugs 4
(Liscovitch-Brauer et al., ATAC CRISPR-cas9 22,84

2021)
(McFarland et al., 2020) RNA drugs, CRISPR-cas9 18
(Mimitou et al., 2021) ATAC + protein CRISPR-cas9 6
(Norman et al., 2019) RNA CRISPRa 237
(Papalexi et al., 2021) RNA + protein CRISPR-cas9 11,99
(Pierce et al., 2021) ATAC CRISPRi 41,41,41
(Replogle et al., 2022) RNA CRISPRi 2058, 2394, 9867
(Schiebinger et al., 2019) RNA cytokines 2,3
(Schraivogel et al., 2020) RNA CRISPR-cas9 3105*, 4115*
(Shifrut et al., 2018) RNA CRISPR-cas9+TCR& 49
(Srivatsan et al., 2020) RNA drugs 5, 8, 189
(Tian et al., 2019) RNA CRISPRi 27
(Tian et al., 2021) RNA CRISPRa, CRISPRi 101, 185
(Weinreb et al., 2020) RNA cytokines 5
(Xie et al., 2017) RNA CRISPR-cas9 229
(Zhao et al., 2021) RNA drugs 7

*: perturbation total treats perturbations A, B, and (A and B) as three unique perturbations
&
: TCR receptor stimulation
18
Methods
scATAC-seq
Data acquisition
We included scATAC-seq data from three different sources: Spear-ATAC (Pierce et
al., 2021), CRISPR-sciATAC (Liscovitch-Brauer et al., 2021), and ASAP-seq
(Mimitou et al., 2019). All data that was used in our analysis can be programmatically
downloaded with scripts that are provided in our code repository
(https://github.com/sanderlab/scPerturb).
scATAC-seq is a biomolecular technique to assess chromatin accessibility within

single cells (Buenrostro et al., 2015; Cusanovich et al., 2015). The starting point of
our data processing pipeline are BED-like tabular fragment files, in which each line
represents a unique ATAC-seq fragment captured by the assay. Each fragment is
mapped to a genomic interval and a cell barcode. The goal of our pipeline is to
extract standardized features from this information. Those are:
● Embeddings derived from Latent-Semantic-Indexing (LSI) (Cusanovich et al.,

2015) with 30 dimensions for each cell (a dimensionality reduction method
that is well-suited for the sparsity of the data)
● Gene scores that measure the chromatin accessibility around each gene for
each cell (the weighted sum of fragment counts around the neighborhood of a
gene’s transcription start site where more distant counts contribute less)
● A peak-barcode matrix that quantifies the chromatin accessibility at (data-set
specific) consensus peaks (genomic intervals) for each cell
● ChromVar scores (Schep et al., 2017), which quantify the activity of a set of
transcription factors for each cell, using transcription factor footprints as
defined in (Vierstra et al., 2020)
● Marker-peaks per perturbation target, quantifying the differential regulation of
highly variable peaks for each type of perturbation
These features were computed using the ArchR framework version 1.0.1 (Granja et
al., 2021) with standard parameters unless otherwise stated. We provide each
feature set as a dedicated h5ad file on scperturb.org, and our analysis roughly
follows the pipeline proposed in Spear-ATAC (Pierce et al., 2021), as detailed below.
Note that these features were originally developed for scATAC-seq data on
non-perturbed cells, with goals such as the identification of cell types, discovery of
cell type-specific regulatory elements, or reconstruction of cellular differentiation
trajectories (Buenrostro et al., 2013; Satpathy et al., 2019).
19
Pre-processing
Filtering out cells of low quality: To ensure a consistent and homogenous quality
throughout the different data sets, we filtered out cells with fewer than 1000 and
more than 100,000 mapped fragments. We further required a minimum transcription
start site enrichment score of 4 to ensure a sufficient signal to noise ratio. See
ArchR’s ‘createArrowFile’ function for details.
For the Spear-ATAC data set we ran ArchR’s getValidBarcodes function on

processed 10x Cell Ranger files to subset the data set to valid barcodes. For the
other datasets these files were unavailable, and we relied on the original authors’
pre-processing of barcodes.
Assigning sgRNAs to barcodes: For the Spear-ATAC and CRISPR_sciATAC

datasets we had access to cell barcode-sgRNA count matrices (see original
publications for details). We assigned the sgRNA with the highest counts to a cell
barcode if the sgRNA count exceeded 20 and if that sgRNA combined at least 80%
of all sgRNA counts. Cells that could not be assigned a sgRNA were left in the data
set. For the ASAP-seq dataset a barcode-sgRNA matrix was not available. Instead,
we relied on a sgRNA assignment downloaded from the study’s GitHub repository
(Lareau, 2021).
Feature computation
All features described in the overview above were computed with ArchR functions.
For details inspect the “fragments2outputs.R” script in our code repository (see Data
Availability).
scRNA-seq
Data acquisition
Datasets were downloaded from public databases following data availability
directions in the source papers. When available from the authors, unnormalized
pre-processed cell-by-gene matrices were used. Supplemental information from the
papers were used in data analysis when applicable.
Data processing
Analysis started from unfiltered, unnormalized cell-by-gene matrices as provided by
source papers. For one dataset, preprocessed cell-by-gene matrices were
unavailable; pre-processing was performed following the procedure outlined in the
original paper, directly using supplied code (Gehring et al., 2020). For datasets with
cell barcodes, barcode assignments for cells were taken from the original paper
when available; when not available, barcode assignment was performed as
described in the methods section of the relevant paper. If multiple guides were
assigned to the same cell, the guides were listed in decreasing order of counts in the
20
final data object. The code used for processing each individual dataset, including
barcode assignment, is available in our code repository.
Datasets were imported into AnnData objects using Scanpy (versions 1.7.2–1.9.1)
(Wolf et al., 2018). Metadata was taken from the original papers when available. For
cell lines, information on sex, age, disease, and origin were taken from Cellosaurus
(Bairoch, 2018). Metadata columns are described in (Supp Table 2). Items listed in
bold are included for all datasets.
Datasets are saved as .h5ad files and as .mtx files with obs and var as separate .csv
files. Code is supplied in our code repository for the import of .mtx files into Seurat.
Data analysis
Before calculating the E-distances (Fig 4), cells and genes were filtered using
Scanpy (versions 1.7.2–1.9.1) (Wolf et al., 2018). All .h5ad objects published on the
resource were saved using Scanpy 1.9.1. Cells were kept if they had a minimum of
1000 UMI counts, and genes with a minimum of 50 cells. 2000 highly variable genes
were selected using scanpy.pp.find_variable_genes with flavor ‘seurat_v3’.
We normalized the count matrix using scanpy.pp.normalize_total and
log-transformed the data using scanpy.pp.log1p; We did not z-scale the data.
Next, we computed PCA based on the highly variable genes. The E-distances were
computed in that PCA space using 50 components and Euclidean distance. To avoid
problems due to different numbers of cells per perturbation, we subsampled each
dataset such that all perturbations had the same number of cells. We removed all
perturbations with fewer than 50 cells and then subsampled to the number of cells in
the smallest perturbation left after filtering. Large parts of our analysis were
parallelized as workflows using snakemake (Mölder et al., 2021).
E-distance
The E-distance is a statistical distance between high-dimensional distributions and
has been used to define a multivariate two-sample test, called the E-test (Rizzo and
Székely, 2016). It is more commonly known as energy distance, stemming from the
original interpretation using gravitational energy in physics. Formally, it
contextualizes the notion that two distributions of points in a high-dimensional space
are distinguishable if they are far apart compared to the width of both distributions
(Fig 3A). More specifically,
Let and be samples from two distributions ,
corresponding to two sets of N and M cells, respectively.
We define
21
and defined accordingly. We used the squared euclidean distance when

calculating cell-wise distances. Intuitively, is the mean distance between cells
from the two distributions, while describes the mean distance between a cell
from to another cell from . The energy distance between and is defined
as:
E-test calculation
The E-test was performed as a Monte Carlo permutation test using the E-distance as
test statistic. For each dataset and each perturbation within that dataset, we took the
cells and combined them with the unperturbed cells. Then, we shuffled the
perturbation labels and computed the E-distance between the two resulting groups.
We repeated this process 100 times. The number of times that this shuffled
E-distance to unperturbed was larger than the unshuffled one divided by 100 yields a
p-value, which we report for almost all datasets in our resource (Supp Table 3). We
corrected for multiple testing using the Holm-Sidak method per dataset.
Cell type hierarchy

Cell type annotations celltype.l2 were used as provided by (Hao et al., 2021).
Doublets were removed and data was subset to 91 cells per remaining cell type,
which is the largest number such that all key cell types had at least that many cells.
After subsetting, the data was processed as for other RNA datasets. Protein data
was CLR normalized using Muon 0.1.2 and log-transformed prior to PCA (Bredikhin
et al., 2022). The hierarchy was computed using
scipy.cluster.hierarchy.linkage from scipy 1.8.0 with method “single”.
Subsampling analysis
At each subsampling point we computed detailed E-statistics (E-distances, delta,
sigma, E-test results) from each perturbation to the corresponding unperturbed cells
of that dataset using PCA with 50 components based on 2000 highly variable genes,
except specified otherwise. We downsampled raw UMI counts using the function
scanpy.pp.downsample_counts on raw counts, then preprocessed (normalized,
log1p-transformed, etc.) the data as previously described. Cells were downsampled
to the same number at each subsampling step across all perturbations to avoid
comparability issues. If possible, we recalculated the PCA while keeping the highly
variable genes originally obtained from the complete dataset. Figures 5C, 5D and
Supplemental Figures 5A, 5B were computed as a running loss of E-test significance
22
(p-value<0.05) of formerly – i.e. prior to any subsampling – significant perturbations

while subsampling, then normalized across datasets through division by the total
number of formally significant perturbations in that datasets.
Advice for single-cell perturbation analysis

Resource users should be aware that memory requirements quickly become a
limiting factor, especially with the newer, larger datasets, such as
ReplogleWeissman2022 with >2.5 million cells across more than 9000 perturbations
(Replogle et al., 2022). For example, the E-distance presented here for calculating
distances between perturbed sets of cells relies on principal component analysis
(PCA), but computing PCA for all data in this dataset was not possible with 500GB of
memory without modifications to accelerate computation. Going forward,
computational methods will need to be modified as in (Dhapola et al., 2022) to
reduce memory load, or datasets will need to be subsampled. Additionally, the .h5ad
datasets shared in this resource can be programmatically accessed using .h5py, and
perturbations of interest extracted without requiring full dataset access.
To our knowledge, there are not yet established best practices for analysis of
single-cell perturbation data. DESeq2 is frequently used for differential expression
testing, as it can be applied to pseudo-bulk profiles of each perturbation (Love et al.,
2014). An optional next step would be enrichment analysis of the resulting genes.
Averaging single-cell measurements over cells per perturbation simplifies analysis
and reduces the effect of measurement noise significantly but comes at the cost of
removing all system-intrinsic biologically relevant information in cell-to-cell variation.
In many studies, these average profiles are then embedded using a dimensionality
reduction method of choice and subsequently clustered to reveal groups of
perturbations with potentially similar targets (Norman et al., 2019; Replogle et al.,
2022).
Funding / Acknowledgements
● National Resource for Network Biology (NRNB, P41GM103504)
● Supported by the Wellcome Leap ∆Tissue Program
● Deutsche Forschungsgemeinschaft (DFG, RTG2424 CompCancer, Beyond
the Exome)
● Einstein Stiftung Berlin (Einstein visiting fellow program)
● Computation was in part performed on the HPC for Research cluster of the
Berlin Institute of Health.
● We appreciate informative conversations with Yuge Ji, helpful code
suggestions from Garrett Wong, and computational support from Aaron
Kollasch. We also appreciate preprint review comment from Arcadia Science’s
preprint review initiative (Gregory P. Way, Natalie Davidson, Erik Serrano,
Parker Hicks, Jenna Tomkinson, Dave Bunten).
23
References
https://raw.githubusercontent.com/caleblareau/asap_reproducibility/master/C
D4_CRISPR_asapseq/output/Signac/after_filter_Signac/HTO_res_filtered.txt
Adamson, B., Norman, T.M., Jost, M., Cho, M.Y., Nuñez, J.K., Chen, Y., Villalta, J.E.,
Gilbert, L.A., Horlbeck, M.A., Hein, M.Y., Pak, R.A., Gray, A.N., Gross, C.A.,
Dixit, A., Parnas, O., Regev, A., Weissman, J.S., 2016. A Multiplexed
Single-Cell CRISPR Screening Platform Enables Systematic Dissection of the
Unfolded Protein Response. Cell 167, 1867-1882.e21.
https://doi.org/10.1016/j.cell.2016.11.048
Aissa, A.F., Islam, A.B.M.M.K., Ariss, M.M., Go, C.C., Rader, A.E., Conrardy, R.D.,
Gajda, A.M., Rubio-Perez, C., Valyi-Nagy, K., Pasquinelli, M., Feldman, L.E.,
Green, S.J., Lopez-Bigas, N., Frolov, M.V., Benevolenskaya, E.V., 2021.
Single-cell transcriptional changes associated with drug tolerance and
response to combination therapies in cancer. Nat. Commun. 12, 1628.
https://doi.org/10.1038/s41467-021-21884-z
Artis, D., Spits, H., 2015. The biology of innate lymphoid cells. Nature 517, 293–301.
https://doi.org/10.1038/nature14189
Bairoch, A., 2018. The Cellosaurus, a Cell-Line Knowledge Resource. J. Biomol.
Tech. JBT 29, 25–38. https://doi.org/10.7171/jbt.18-2902-002
Barrett, T., Wilhite, S.E., Ledoux, P., Evangelista, C., Kim, I.F., Tomashevsky, M.,
Marshall, K.A., Phillippy, K.H., Sherman, P.M., Holko, M., Yefanov, A., Lee, H.,
Zhang, N., Robertson, C.L., Serova, N., Davis, S., Soboleva, A., 2013. NCBI
GEO: archive for functional genomics data sets—update. Nucleic Acids Res.
41, D991–D995. https://doi.org/10.1093/nar/gks1193
Bertin, P., Rector-Brooks, J., Sharma, D., Gaudelet, T., Anighoro, A., Gross, T.,
Martinez-Pena, F., Tang, E.L., S, S.M., Regep, C., Hayter, J., Korablyov, M.,
Valiante, N., van der Sloot, A., Tyers, M., Roberts, C., Bronstein, M.M.,
Lairson, L.L., Taylor-King, J.P., Bengio, Y., 2022. RECOVER: sequential
model optimization platform for combination drug repurposing identifies novel
synergistic compounds in vitro. https://doi.org/10.48550/arXiv.2202.04202
Bredikhin, D., Kats, I., Stegle, O., 2022. MUON: multimodal omics analysis
framework. Genome Biol. 23, 42. https://doi.org/10.1186/s13059-021-02577-8
Broad Institute, 2022. Single Cell Portal [WWW Document]. URL
https://singlecell.broadinstitute.org/single_cell (accessed 8.17.22).
Buenrostro, J.D., Giresi, P.G., Zaba, L.C., Chang, H.Y., Greenleaf, W.J., 2013.
Transposition of native chromatin for fast and sensitive epigenomic profiling of
open chromatin, DNA-binding proteins and nucleosome position. Nat.
Methods 10, 1213–1218. https://doi.org/10.1038/nmeth.2688
Buenrostro, J.D., Wu, B., Litzenburger, U.M., Ruff, D., Gonzales, M.L., Snyder, M.P.,
Chang, H.Y., Greenleaf, W.J., 2015. Single-cell chromatin accessibility reveals
principles of regulatory variation. Nature 523, 486–490.
https://doi.org/10.1038/nature14590
Burkhardt, D.B., Stanley, J.S., Tong, A., Perdigoto, A.L., Gigante, S.A., Herold, K.C.,
Wolf, G., Giraldez, A.J., van Dijk, D., Krishnaswamy, S., 2021. Quantifying the
effect of experimental perturbations at single-cell resolution. Nat. Biotechnol.
24
39, 619–629. https://doi.org/10.1038/s41587-020-00803-5

Chang, M.T., Shanahan, F., Nguyen, T.T.T., Staben, S.T., Gazzard, L., Yamazoe, S.,
Wertz, I.E., Piskol, R., Yang, Y.A., Modrusan, Z., Haley, B., Evangelista, M.,
Malek, S., Foster, S.A., Ye, X., 2022. Identifying transcriptional programs
underlying cancer drug response with TraCe-seq. Nat. Biotechnol. 40, 86–93.
https://doi.org/10.1038/s41587-021-01005-3
Chen, H., Lareau, C., Andreani, T., Vinyard, M.E., Garcia, S.P., Clement, K.,
Andrade-Navarro, M.A., Buenrostro, J.D., Pinello, L., 2019. Assessment of
computational methods for the analysis of single-cell ATAC-seq data. Genome
Biol. 20, 241. https://doi.org/10.1186/s13059-019-1854-5
Chen, W.S., Zivanovic, N., van Dijk, D., Wolf, G., Bodenmiller, B., Krishnaswamy, S.,
2020. Uncovering axes of variation among single-cell cancer specimens. Nat.
Methods 17, 302–310. https://doi.org/10.1038/s41592-019-0689-z
Cusanovich, D.A., Daza, R., Adey, A., Pliner, H.A., Christiansen, L., Gunderson,
K.L., Steemers, F.J., Trapnell, C., Shendure, J., 2015. Multiplex single-cell
profiling of chromatin accessibility by combinatorial cellular indexing. Science
348, 910–914. https://doi.org/10.1126/science.aab1601
Dann, E., Henderson, N.C., Teichmann, S.A., Morgan, M.D., Marioni, J.C., 2022.
Differential abundance testing on single-cell data using k-nearest neighbor
graphs. Nat. Biotechnol. 40, 245–253.
Datlinger, P., Rendeiro, A.F., Boenke, T., Senekowitsch, M., Krausgruber, T., Barreca,
D., Bock, C., 2021. Ultra-high-throughput single-cell RNA sequencing and
perturbation screening with combinatorial fluidic indexing. Nat. Methods 18,
635–642. https://doi.org/10.1038/s41592-021-01153-z
Datlinger, P., Rendeiro, A.F., Schmidl, C., Krausgruber, T., Traxler, P., Klughammer,
J., Schuster, L.C., Kuchler, A., Alpar, D., Bock, C., 2017. Pooled CRISPR
screening with single-cell transcriptome readout. Nat. Methods 14, 297–301.
https://doi.org/10.1038/nmeth.4177
Dhapola, P., Rodhe, J., Olofzon, R., Bonald, T., Erlandsson, E., Soneji, S., Karlsson,
G., 2022. Scarf enables a highly memory-efficient analysis of large-scale
single-cell genomics data. Nat. Commun. 13, 4616.
Diehl, A.D., Meehan, T.F., Bradford, Y.M., Brush, M.H., Dahdul, W.M., Dougall, D.S.,
He, Y., Osumi-Sutherland, D., Ruttenberg, A., Sarntivijai, S., Van Slyke, C.E.,
Vasilevsky, N.A., Haendel, M.A., Blake, J.A., Mungall, C.J., 2016. The Cell
Ontology 2016: enhanced content, modularization, and ontology
interoperability. J. Biomed. Semant. 7, 44.
Dixit, A., Parnas, O., Li, B., Chen, J., 2016. Perturb-Seq: Dissecting Molecular
Circuits with Scalable Single-Cell RNA Profiling of Pooled Genetic Screens.
Cell 167.
Duan, B., Zhou, C., Zhu, C., Yu, Y., Li, G., Zhang, S., Zhang, C., Ye, X., Ma, H., Qu,
S., Zhang, Z., Wang, P., Sun, S., Liu, Q., 2019. Model-based understanding of
single-cell CRISPR screening. Nat. Commun. 10, 2233.
https://doi.org/10.1038/s41467-019-10216-x
Fischer, D.S., Dony, L., König, M., Moeed, A., Zappia, L., Heumos, L., Tritschler, S.,
Holmberg, O., Aliee, H., Theis, F.J., 2021. Sfaira accelerates data and model
reuse in single cell genomics. Genome Biol. 22, 248.
25
Forcato, M., Romano, O., Bicciato, S., 2021. Computational methods for the
integrative analysis of single-cell data. Brief. Bioinform. 22.
https://doi.org/10.1093/bib/bbaa042
Frangieh, C.J., Melms, J.C., Thakore, P.I., Geiger-Schuller, K.R., Ho, P., Luoma,
A.M., Cleary, B., Jerby-Arnon, L., Malu, S., Cuoco, M.S., Zhao, M., Ager, C.R.,
Rogava, M., Hovey, L., Rotem, A., Bernatchez, C., Wucherpfennig, K.W.,
Johnson, B.E., Rozenblatt-Rosen, O., Schadendorf, D., Regev, A., Izar, B.,
2021. Multimodal pooled Perturb-CITE-seq screens in patient models define
mechanisms of cancer immune evasion. Nat. Genet. 53, 332–341.
Franz, A., Coscia, F., Shen, C., Charaoui, L., Mann, M., Sander, C., 2021. Molecular
response to PARP1 inhibition in ovarian cancer cells as determined by mass
spectrometry based proteomics. J. Ovarian Res. 14, 140.
Gasperini, M., Hill, A.J., McFaline-Figueroa, J.L., Martin, B., Kim, S., Zhang, M.D.,
Jackson, D., Leith, A., Schreiber, J., Noble, W.S., Trapnell, C., Ahituv, N.,
Shendure, J., 2019. A Genome-wide Framework for Mapping Gene
Regulation via Cellular Genetic Screens. Cell 176, 377-390.e19.
Gatto, L., Aebersold, R., Cox, J., Demichev, V., Derks, J., Emmott, E., Franks, A.M.,
Ivanov, A.R., Kelly, R.T., Khoury, L., Leduc, A., MacCoss, M.J., Nemes, P.,
Perlman, D.H., Petelski, A.A., Rose, C.M., Schoof, E.M., Van Eyk, J.,
Vanderaa, C., Yates III, J.R., Slavov, N., 2022. Initial recommendations for
performing, benchmarking, and reporting single-cell proteomics experiments.
https://doi.org/10.48550/arXiv.2207.10815
Gehring, J., Hwee Park, J., Chen, S., Thomson, M., Pachter, L., 2020. Highly
multiplexed single-cell RNA-seq by DNA oligonucleotide tagging of cellular
proteins. Nat. Biotechnol. 38, 35–38.
Gilbert, L.A., Horlbeck, M.A., Adamson, B., Villalta, J.E., Chen, Y., Whitehead, E.H.,
Guimaraes, C., Panning, B., Ploegh, H.L., Bassik, M.C., Qi, L.S., Kampmann,
M., Weissman, J.S., 2014. Genome-Scale CRISPR-Mediated Control of Gene
Repression and Activation. Cell 159, 647–661.
Granja, J.M., Corces, M.R., Pierce, S.E., Bagdatli, S.T., Choudhry, H., Chang, H.Y.,
Greenleaf, W.J., 2021. ArchR is a scalable software package for integrative
single-cell chromatin accessibility analysis. Nat. Genet. 53, 403–411.
Gross, T., Blüthgen, N., 2020. Identifiability and experimental design in perturbation
studies. Bioinformatics 36, i482–i489.
https://doi.org/10.1093/bioinformatics/btaa404
Gross, T., Wongchenko, M.J., Yan, Y., Blüthgen, N., 2019. Robust network inference
using response logic. Bioinformatics 35, i634–i642.
https://doi.org/10.1093/bioinformatics/btz326
Hao, Y., Hao, S., Andersen-Nissen, E., Mauck, W.M., Zheng, S., Butler, A., Lee,
M.J., Wilk, A.J., Darby, C., Zager, M., Hoffman, P., Stoeckius, M., Papalexi, E.,
Mimitou, E.P., Jain, J., Srivastava, A., Stuart, T., Fleming, L.M., Yeung, B.,
Rogers, A.J., McElrath, J.M., Blish, C.A., Gottardo, R., Smibert, P., Satija, R.,
2021. Integrated analysis of multimodal single-cell data. Cell 184,
3573-3587.e29. https://doi.org/10.1016/j.cell.2021.04.048
26
Haque, A., Engel, J., Teichmann, S.A., Lönnberg, T., 2017. A practical guide to
single-cell RNA-sequencing for biomedical research and clinical applications.
Genome Med. 9, 75. https://doi.org/10.1186/s13073-017-0467-4
Jaitin, D.A., Weiner, A., Yofe, I., Lara-Astiaso, D., Keren-Shaul, H., David, E.,
Salame, T.M., Tanay, A., Oudenaarden, A. van, Amit, I., 2016. Dissecting
Immune Circuits by Linking CRISPR-Pooled Screens with Single-Cell
RNA-Seq. Cell 167, 1883-1896.e15. https://doi.org/10.1016/j.cell.2016.11.039
Ji, Y., Lotfollahi, M., Wolf, F.A., Theis, F.J., 2021. Machine learning for perturbational
single-cell omics. Cell Syst. 12, 522–537.
https://doi.org/10.1016/j.cels.2021.05.016
Jin, K., Schnell, D., Li, G., Salomonis, N., Prasath, V.B.S., Szczesniak, R., Aronow,
B.J., 2022. CellDrift: Inferring Perturbation Responses in Temporally-Sampled
Single Cell Data. https://doi.org/10.1101/2022.04.13.488194
Kharchenko, P.V., 2021. The triumphs and limitations of computational methods for
scRNA-seq. Nat. Methods 18, 723–732.
Lance, C., Luecken, M.D., Burkhardt, D.B., Cannoodt, R., Rautenstrauch, P.,
Laddach, A., Ubingazhibov, A., Cao, Z.-J., Deng, K., Khan, S., Liu, Q.,
Russkikh, N., Ryazantsev, G., Ohler, U., Participants, N. 2021 M. data
integration competition, Pisco, A.O., Bloom, J., Krishnaswamy, S., Theis, F.J.,
2022. Multimodal single cell data integration challenge: results and lessons
learned. https://doi.org/10.1101/2022.04.11.487796
Lareau, Caleb.A., 2021. asap_reproducibility.
Liscovitch-Brauer, N., Montalbano, A., Deng, J., Méndez-Mancilla, A., Wessels,
H.-H., Moss, N.G., Kung, C.-Y., Sookdeo, A., Guo, X., Geller, E., Jaini, S.,
Smibert, P., Sanjana, N.E., 2021. Profiling the genetic determinants of
chromatin accessibility with scalable single-cell CRISPR screens. Nat.
Biotechnol. 39, 1270–1277. https://doi.org/10.1038/s41587-021-00902-x
Lotfollahi, M., Naghipourfar, M., Theis, F.J., Wolf, F.A., 2020. Conditional
out-of-distribution generation for unpaired data using transfer VAE.
Bioinformatics 36, i610–i617. https://doi.org/10.1093/bioinformatics/btaa800
Lotfollahi, M., Wolf, F.A., Theis, F.J., 2019. scGen predicts single-cell perturbation
responses. Nat. Methods 16, 715–721.
Love, M.I., Huber, W., Anders, S., 2014. Moderated estimation of fold change and
dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550.
Luecken, M.D., Büttner, M., Chaichoompu, K., Danese, A., Interlandi, M., Mueller,
M.F., Strobl, D.C., Zappia, L., Dugas, M., Colomé-Tatché, M., Theis, F.J.,
2022. Benchmarking atlas-level data integration in single-cell genomics. Nat.
Methods 19, 41–50. https://doi.org/10.1038/s41592-021-01336-8
Luecken, M.D., Theis, F.J., 2019. Current best practices in single-cell RNA-seq
analysis: a tutorial. Mol. Syst. Biol. 15, e8746.
https://doi.org/10.15252/msb.20188746
McFarland, J.M., Paolella, B.R., Warren, A., Geiger-Schuller, K., Shibue, T.,
Rothberg, M., Kuksenko, O., Colgan, W.N., Jones, A., Chambers, E., Dionne,
D., Bender, S., Wolpin, B.M., Ghandi, M., Tirosh, I., Rozenblatt-Rosen, O.,
Roth, J.A., Golub, T.R., Regev, A., Aguirre, A.J., Vazquez, F., Tsherniak, A.,
2020. Multiplexed single-cell transcriptional response profiling to define
cancer vulnerabilities and therapeutic mechanism of action. Nat. Commun. 11,
27
4296. https://doi.org/10.1038/s41467-020-17440-w
Mimitou, E.P., Cheng, A., Montalbano, A., Hao, S., Stoeckius, M., Legut, M., Roush,
T., Herrera, A., Papalexi, E., Ouyang, Z., Satija, R., Sanjana, N.E., Koralov,
S.B., Smibert, P., 2019. Multiplexed detection of proteins, transcriptomes,
clonotypes and CRISPR perturbations in single cells. Nat. Methods 16,
409–412. https://doi.org/10.1038/s41592-019-0392-0
Mimitou, E.P., Lareau, C.A., Chen, K.Y., Zorzetto-Fernandes, A.L., Hao, Y.,
Takeshima, Y., Luo, W., Huang, T.-S., Yeung, B.Z., Papalexi, E., Thakore, P.I.,
Kibayashi, T., Wing, J.B., Hata, M., Satija, R., Nazor, K.L., Sakaguchi, S.,
Ludwig, L.S., Sankaran, V.G., Regev, A., Smibert, P., 2021. Scalable,
multimodal profiling of chromatin accessibility, gene expression and protein
levels in single cells. Nat. Biotechnol. 39, 1246–1258.
Mölder, F., Jablonski, K.P., Letcher, B., Hall, M.B., Tomkins-Tinch, C.H., Sochat, V.,
Forster, J., Lee, S., Twardziok, S.O., Kanitz, A., Wilm, A., Holtgrewe, M.,
Rahmann, S., Nahnsen, S., Köster, J., 2021. Sustainable data analysis with
Snakemake. https://doi.org/10.12688/f1000research.29032.2
Norman, T.M., Horlbeck, M.A., Replogle, J.M., Ge, A.Y., Xu, A., Jost, M., Gilbert,
L.A., Weissman, J.S., 2019. Exploring genetic interaction manifolds
constructed from rich single-cell phenotypes. Science 365, 786–793.
https://doi.org/10.1126/science.aax4438
Papalexi, E., Mimitou, E.P., Butler, A.W., Foster, S., Bracken, B., Mauck, W.M.,
Wessels, H.-H., Hao, Y., Yeung, B.Z., Smibert, P., Satija, R., 2021.
Characterizing the molecular regulation of inhibitory immune checkpoints with
multimodal single-cell screens. Nat. Genet. 53, 322–331.
Pierce, S.E., Granja, J.M., Greenleaf, W.J., 2021. High-throughput single-cell
chromatin accessibility CRISPR screens enable unbiased identification of
regulatory networks in cancer. Nat. Commun. 12, 2969.
https://doi.org/10.1038/s41467-021-23213-w
Pratapa, A., Jalihal, A.P., Law, J.N., Bharadwaj, A., Murali, T.M., 2020.
Benchmarking algorithms for gene regulatory network inference from
single-cell transcriptomic data. Nat. Methods 17, 147–154.
Preuer, K., Lewis, R.P.I., Hochreiter, S., Bender, A., Bulusu, K.C., Klambauer, G.,
2018. DeepSynergy: predicting anti-cancer drug synergy with Deep Learning.
Bioinformatics 34, 1538–1546. https://doi.org/10.1093/bioinformatics/btx806
Przybyla, L., Gilbert, L.A., 2022. A new era in functional genomics screens. Nat. Rev.
Genet. 23, 89–103. https://doi.org/10.1038/s41576-021-00409-w
Replogle, J.M., Saunders, R.A., Pogson, A.N., Hussmann, J.A., Lenail, A., Guna, A.,
Mascibroda, L., Wagner, E.J., Adelman, K., Lithwick-Yanai, G., Iremadze, N.,
Oberstrass, F., Lipson, D., Bonnar, J.L., Jost, M., Norman, T.M., Weissman,
J.S., 2022. Mapping information-rich genotype-phenotype landscapes with
genome-scale Perturb-seq. Cell 185, 2559-2575.e28.
Rizzo, M.L., Székely, G.J., 2016. Energy distance. WIREs Comput. Stat. 8, 27–38.
https://doi.org/10.1002/wics.1375
Rubin, A.J., Parker, K.R., Satpathy, A.T., Qi, Y., Wu, B., Ong, A.J., Mumbach, M.R.,
Ji, A.L., Kim, D.S., Cho, S.W., Zarnegar, B.J., Greenleaf, W.J., Chang, H.Y.,
Khavari, P.A., 2019. Coupled Single-Cell CRISPR Screening and Epigenomic
28
Profiling Reveals Causal Gene Regulatory Networks. Cell 176, 361-376.e17.

Satpathy, A.T., Granja, J.M., Yost, K.E., Qi, Y., Meschi, F., McDermott, G.P., Olsen,
B.N., Mumbach, M.R., Pierce, S.E., Corces, M.R., Shah, P., Bell, J.C., Jhutty,
D., Nemec, C.M., Wang, J., Wang, L., Yin, Y., Giresi, P.G., Chang, A.L.S.,
Zheng, G.X.Y., Greenleaf, W.J., Chang, H.Y., 2019. Massively parallel
single-cell chromatin landscapes of human immune cell development and
intratumoral T cell exhaustion. Nat. Biotechnol. 37, 925–936.
Schep, A.N., Wu, B., Buenrostro, J.D., Greenleaf, W.J., 2017. chromVAR: inferring
transcription-factor-associated accessibility from single-cell epigenomic data.
Nat. Methods 14, 975–978. https://doi.org/10.1038/nmeth.4401
Schiebinger, G., Shu, J., Tabaka, M., Cleary, B., Subramanian, V., Solomon, A.,
Gould, J., Liu, S., Lin, S., Berube, P., Lee, L., Chen, J., Brumbaugh, J.,
Rigollet, P., Hochedlinger, K., Jaenisch, R., Regev, A., Lander, E.S., 2019.
Optimal-Transport Analysis of Single-Cell Gene Expression Identifies
Developmental Trajectories in Reprogramming. Cell 176, 928-943.e22.
Schraivogel, D., Gschwind, A.R., Milbank, J.H., Leonce, D.R., Jakob, P., Mathur, L.,
Korbel, J.O., Merten, C.A., Velten, L., Steinmetz, L.M., 2020. Targeted
Perturb-seq enables genome-scale genetic screens in single cells. Nat.
Methods 17, 629–635. https://doi.org/10.1038/s41592-020-0837-5
Shifrut, E., Carnevale, J., Tobin, V., Roth, T.L., Woo, J.M., Bui, C.T., Li, P.J., Diolaiti,
M.E., Ashworth, A., Marson, A., 2018. Genome-wide CRISPR Screens in
Primary Human T Cells Reveal Key Regulators of Immune Function. Cell 175,
1958-1971.e15. https://doi.org/10.1016/j.cell.2018.10.024
Srivatsan, S.R., McFaline-Figueroa, J.L., Ramani, V., Saunders, L., Cao, J., Packer,
J., Pliner, H.A., Jackson, D.L., Daza, R.M., Christiansen, L., Zhang, F.,
Steemers, F., Shendure, J., Trapnell, C., 2020. Massively multiplex chemical
transcriptomics at single-cell resolution. Science 367, 45–51.
https://doi.org/10.1126/science.aax6234
Stathias, V., Turner, J., Koleti, A., Vidovic, D., Cooper, D., Fazel-Najafabadi, M.,
Pilarczyk, M., Terryn, R., Chung, C., Umeano, A., Clarke, D.J.B., Lachmann,
A., Evangelista, J.E., Ma’ayan, A., Medvedovic, M., Schürer, S.C., 2020.
LINCS Data Portal 2.0: next generation access point for
perturbation-response signatures. Nucleic Acids Res. 48, D431–D439.
https://doi.org/10.1093/nar/gkz1023
Svensson, V., da Veiga Beltrame, E., Pachter, L., 2020. A curated database reveals
trends in single-cell transcriptomics. Database 2020.
https://doi.org/10.1093/database/baaa073
Székely, G.J., Rizzo, M.L., 2013. Energy statistics: A class of statistics based on
distances. J. Stat. Plan. Inference 143, 1249–1272.
https://doi.org/10.1016/j.jspi.2013.03.018
Tian, R., Abarientos, A., Hong, J., Hashemi, S.H., Yan, R., Dräger, N., Leng, K.,
Nalls, M.A., Singleton, A.B., Xu, K., Faghri, F., Kampmann, M., 2021.
Genome-wide CRISPRi/a screens in human neurons link lysosomal failure to
ferroptosis. Nat. Neurosci. 24, 1020–1034.
Tian, R., Gachechiladze, M.A., Ludwig, C.H., Laurie, M.T., Hong, J.Y., Nathaniel, D.,
Prabhu, A.V., Fernandopulle, M.S., Patel, R., Abshari, M., Ward, M.E.,
29
Kampmann, M., 2019. CRISPR Interference-Based Platform for Multimodal

Genetic Screens in Human iPSC-Derived Neurons. Neuron 104, 239-255.e12.
https://doi.org/10.1016/j.neuron.2019.07.014
Tsherniak, A., Vazquez, F., Montgomery, P.G., Weir, B.A., Kryukov, G., Cowley, G.S.,
Gill, S., Harrington, W.F., Pantel, S., Krill-Burger, J.M., Meyers, R.M., Ali, L.,
Goodale, A., Lee, Y., Jiang, G., Hsiao, J., Gerath, W.F.J., Howell, S., Merkel,
E., Ghandi, M., Garraway, L.A., Root, D.E., Golub, T.R., Boehm, J.S., Hahn,
W.C., 2017. Defining a Cancer Dependency Map. Cell 170, 564-576.e16.
Vierstra, J., Lazar, J., Sandstrom, R., Halow, J., Lee, K., Bates, D., Diegel, M., Dunn,
D., Neri, F., Haugen, E., Rynes, E., Reynolds, A., Nelson, J., Johnson, A.,
Frerker, M., Buckley, M., Kaul, R., Meuleman, W., Stamatoyannopoulos, J.A.,
2020. Global reference mapping of human transcription factor footprints.
Nature 583, 729–736. https://doi.org/10.1038/s41586-020-2528-x
Vivier, E., Artis, D., Colonna, M., Diefenbach, A., Santo, J.P.D., Eberl, G., Koyasu, S.,
Locksley, R.M., McKenzie, A.N.J., Mebius, R.E., Powrie, F., Spits, H., 2018.
Innate Lymphoid Cells: 10 Years On. Cell 174, 1054–1066.
Weinreb, C., Rodriguez-Fraticelli, A., Camargo, F.D., Klein, A.M., 2020. Lineage
tracing on transcriptional landscapes links state to fate during differentiation.
Science 367. https://doi.org/10.1126/science.aaw3381
Wessels, H.-H., Méndez-Mancilla, A., Papalexi, E., Mauck, W.M., Lu, L., Morris, J.A.,
Mimitou, E., Smibert, P., Sanjana, N.E., Satija, R., 2022. Efficient
combinatorial targeting of RNA transcripts in single cells with Cas13 RNA
Perturb-seq (preprint). Genomics. https://doi.org/10.1101/2022.02.02.478894
Wolf, F.A., Angerer, P., Theis, F.J., 2018. SCANPY: large-scale single-cell gene
expression data analysis. Genome Biol. 19, 15.
Xie, S., Duan, J., Li, B., Zhou, P., Hon, G.C., 2017. Multiplexed Engineering and
Analysis of Combinatorial Enhancer Activity in Single Cells. Mol. Cell 66,
285-299.e5. https://doi.org/10.1016/j.molcel.2017.03.007
Zhao, W., Dovas, A., Spinazzi, E.F., Levitin, H.M., Banu, M.A., Upadhyayula, P.,
Sudhakar, T., Marie, T., Otten, M.L., Sisti, M.B., Bruce, J.N., Canoll, P., Sims,
P.A., 2021. Deconvolution of cell type-specific drug responses in human tumor
tissue with single-cell RNA-seq. Genome Med. 13, 82.
https://doi.org/10.1186/s13073-021-00894-y
30
Supplement
Supp Fig 1: Exemplary information provided for each scPerturb dataset (here for (Norman et
al., 2019)) (A) Number of genes that are detected with at least one count in a cell across all cells. (B)
Total number of UMI counts per cell across all cells. Together with Supp Fig 1A this provides an
overview over both sparsity and quality of the dataset. (C) Number of cells per perturbation.
Depending on the application, perturbations with few cells can be filtered out before down-stream
analysis. High imbalance in cell numbers per perturbation may also lead to biases in models. (D)
Number of cells which received none, a single one, or two perturbations.
31
Supp Fig 2A: Number of cells per dataset by submission date. There is a rapid increase in
published single-cell perturbation datasets around 2019. We speculate that the slight decrease of
dataset numbers after 2021 suggested by the plot is due to the ongoing impact of reduced research in
the earlier phases of the COVID-19 pandemic.
Supp Fig 2B: Harmonization and analysis workflow. Perturbation datasets with single-cell
molecular profiles with at least two perturbations and one control condition (e.g. unperturbed) of
various modality types were identified in a literature search. Data was obtained from public
repositories, and metadata (such as guide identity) from paper supplements. Datasets were
reprocessed to standardize annotations and analyzed in parallel. All datasets are now available for
download from scperturb.org, along with visualizations and summarizing information
32
Supp Fig 3: E-distance extended plots. (A) E-distances between all pairs of perturbations in the
dataset NormanWeissman2019. The color scale is clipped at 5% highest and lowest percentiles.
Clusters of similar perturbations are visible, e.g. a cluster of strongly acting perturbations targeting
CEBPA at the top. (B) Percentage of perturbations with significant E-test to unperturbed cells in each
dataset plotted against the total number of perturbations in the dataset (both in log scale).
33
Supp Fig 5: Tests on robustness of E-statistics to dataset properties and parameters. (A) Effect
of using different numbers of principal components (PCs) from PCA on the number of perturbations
with significant E-test w.r.t. unperturbed cells. The SchraivogelSteinmetz2020 dataset is TAP-seq,
thus has much less genes measured than all other datasets. The faster decrease in significance
observed in this dataset indicates stronger sensitivity on the number of PCs with fewer features
available. (B) Effect of using different numbers of highly variable genes (HVGs) for the PCA
calculation prior to E-testing. For most datasets, E-test results appear to stay comparable between
500 and 4000 HVGs. (C) E-distance computed in a single, joint PCA compared to E-distance
computed in a separate PCA per perturbed-unperturbed combination across three exemplary
datasets. Consistently high Pearson correlations indicate strong equivalence between both
approaches across datasets.
34

SC Perthub Single Cell Omics

Uploaded by

Copyright:

Available Formats

SC Perthub Single Cell Omics

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

SC Perthub Single Cell Omics

Uploaded by

Copyright:

Available Formats

bioRxiv preprint doi: https://doi.org/10.1101/2022.08.20.504663; this version posted January 25, 2023.

The copyright holder for this

scPerturb: Harmonized Single-Cell Perturbation Data

1: Institute of Pathology, Charité – Universitätsmedizin Berlin, Corporate Member of Freie Universität

[Uses of single-cell perturbation data]

[Motivation for a distance measure for high-dimensional expression profiles]

[Motivation for unifying datasets]

Fig 1: Perturbation-response profiling for single cells. Different perturbations act at

these experiments to be scaled up massively. While 32 datasets measure

[Choice of features in scATAC-seq data]

[Count statistics in scRNA-seq]

[Cells per perturbation]

Fig 3: E-statistics describe distinctiveness of perturbations in single-cell data. (A)

Replogle et al., we compute the E-distance after PCA (Principal Component

[E-distance: distinguishing cell types]

[E-distance: perturbation dataset example]

[E-distance: perturbation-perturbation distances]

[E-distance: Effect of number of cells on the E-statistics]

E-distance, as a V-statistic, is a biased estimator (Székely and Rizzo, 2013). Despite

[E-distance: Effect of number of UMI counts on the E-statistics]

[E-distance: robust to calculation choices]

[Overlap of perturbations across studies]

frequently differ as well, and comparisons are further complicated by distinct

[Additional considerations for single cell perturbation experimental design]

(Adamson et al., 2016) RNA CRISPRi 9, 20, 114

(Aissa et al., 2021) RNA drugs 4

(Chang et al., 2022) RNA drugs 4

(Datlinger et al., 2017) RNA CRISPR-cas9+TCR& 97

(Datlinger et al., 2021) RNA CRISPR-cas9+TCR& 48

(Dixit et al., 2016) RNA CRISPR-cas9 31

(Frangieh et al., 2021) RNA + protein CRISPR-cas9 249

(Gasperini et al., 2019) RNA CRISPRi 43314*, 39087*, 16531*

(Gehring et al., 2020) RNA drugs 4

(Liscovitch-Brauer et al., ATAC CRISPR-cas9 22,84

(McFarland et al., 2020) RNA drugs, CRISPR-cas9 18

(Mimitou et al., 2021) ATAC + protein CRISPR-cas9 6

(Norman et al., 2019) RNA CRISPRa 237

(Papalexi et al., 2021) RNA + protein CRISPR-cas9 11,99

(Pierce et al., 2021) ATAC CRISPRi 41,41,41

(Replogle et al., 2022) RNA CRISPRi 2058, 2394, 9867

(Schiebinger et al., 2019) RNA cytokines 2,3

(Schraivogel et al., 2020) RNA CRISPR-cas9 3105*, 4115*

(Shifrut et al., 2018) RNA CRISPR-cas9+TCR& 49

(Srivatsan et al., 2020) RNA drugs 5, 8, 189

(Tian et al., 2019) RNA CRISPRi 27

(Tian et al., 2021) RNA CRISPRa, CRISPRi 101, 185

(Weinreb et al., 2020) RNA cytokines 5

(Xie et al., 2017) RNA CRISPR-cas9 229

(Zhao et al., 2021) RNA drugs 7

scATAC-seq is a biomolecular technique to assess chromatin accessibility within

● Embeddings derived from Latent-Semantic-Indexing (LSI) (Cusanovich et al.,

For the Spear-ATAC data set we ran ArchR’s getValidBarcodes function on

Assigning sgRNAs to barcodes: For the Spear-ATAC and CRISPR_sciATAC

and defined accordingly. We used the squared euclidean distance when

Cell type hierarchy

(p-value<0.05) of formerly – i.e. prior to any subsampling – significant perturbations

Advice for single-cell perturbation analysis

39, 619–629. https://doi.org/10.1038/s41587-020-00803-5

Profiling Reveals Causal Gene Regulatory Networks. Cell 176, 361-376.e17.

(Gasperini et al., 2019) RNA CRISPRi 43314, 39087, 16531*

(Schraivogel et al., 2020) RNA CRISPR-cas9 3105, 4115